Fast VP9-to-AV1 Transcoding based on Block Partitioning Inheritance

Alex Borges, Daniel Palomino, Bruno Zatt, Marcelo Porto, Guilherme Correa Video Technology Research Group (ViTech) Graduate Program in Computing (PPGC), Federal University of Pelotas (UFPel), Brazil {amborges, dpalomino, zatt, porto, gcorrea}@inf.ufpel.edu.br

Abstract— This paper proposes a fast VP9-to-AV1 video compared to VP9 (considering the same image quality). This transcoding algorithm based on block partitioning inheritance. represents an average superiority of 22% for AV1 over VP9, The proposed algorithm relies on the reuse of VP9 block an economy of 1/5 in storage resources and other costs partitioning during the AV1 re-encoding process. This way, the involving video transmission services. exhaustive search for the best block size option is avoided to save encoding time. The reuse of VP9 block partitioning is proposed Video transcoding is the process that converts from one based on a statiscal analysis that shows the relation of block video bitstream format to the same format with different parititioning sizes between VP9 and AV1. The analysis configurations (homogeneous transcoding) or to another demontrates that there is a high probability of the AV1 encoding bitstream format (heterogeneous transcoding), as presented in process to choose block sizes of the same size as in the VP9 Fig. 1. With the advent of AV1, converting legacy content encoding. Experimental results show that the proposed from previous formats, such as VP9, becomes an essential task algorithm is able to accelerate the VP9-to-AV1 transcoding for service providers that intend to benefit from its process by 28% on average at the cost of only 4% increase in the compression efficiency. However, as the computational cost BD-Rate when compared with the complete decoding and re- required by libaom is too high, speeding up the encoding encoding process. process is important to allow fast transcoding without decreasing compression efficiency significantly. Keywords—AV1, VP9, transcoding, video coding VP9 and AV1 share several characteristics that can be I. INTRODUCTION harnessed during the transcoding process. Both VP9 and AV1 In 2015, the Alliance for Open Media (AOMedia) follow a block-based hybrid video coding process, as they consortium was created to develop modern royalty-free video divide the frame into smaller parts called blocks for coding formats for online applications, such as on-demand prediction, transform, and quantization. Blocks can assume video transmission, videoconferences, and live streaming. different sizes and shapes, as defined by the codec Partially based on the VP9 [1], Daala [2] and Thor [3] codecs, specification. To achieve the best compression efficiency, the AOMedia launched the AOMedia Video 1 (AV1) [4] format. encoder needs to find the best block size to use in each region Along with the specification, the libaom [5] reference of the video. Usually, this is performed by exhaustive search software was released in 2018. Since then, many other fast over all block size possibilities, which requires a huge AV1 codecs versions have been developed and released by computational effort. In [13], a method for inheriting the best AOMedia members, such as the Scalable Video Technology block size from H.264/AVC to H.265/HEVC during the for AV1 Encoder (SVT-AV1) [6] developed by Intel and transcoding process is proposed. In [14], the H.265/HEVC Netflix, the CISCO-AV1 [7] developed by CISCO, and the Coding Unit depth information is used to accelerate the rav1e [8] developed by XIPH. transcoding from H.265/HEVC to AV1. Although the methods proposed in [13] and [14] present good results and One of the main goals of AV1 is to overcome the are based on block partitioning inheritance, they focus on the compression efficiency achieved by VP9 and replace it as H.264/AVC and H.265/HEVC standards, which employ a current state-of-the-art technology based on royalty-free very different set of block sizes, partitioning modes and codecs. To accomplish that, AV1 includes several new tools coding tree structure in comparison to VP9. Thus, they cannot and features with much more efficient signal processing be directly employed to accelerate the VP9-to-AV1 operations and frame partitioning structures in comparison to transcoding process. To the best of the authors’ knowledge, VP9. However, this efficiency is achieved at the cost of a there is no other work focusing on VP9-to-AV1 transcoding. considerable complexity increase in comparison to VP9. The authors in [9] and [10] show that the reference libaom codec This paper proposes a fast VP9-to-AV1 transcoding requires an encoding time more than 100 times larger in process based on Block Partitioning Inheritance. The comparison to VP9. Thus, time-saving strategies for AV1, proposed solution saves time by reusing the VP9 block especially those leading to small or no penalties in terms of partitioning direction to filter out AV1 partitioning compression efficiency, are currently essential to reduce this possibilities during the re-encoding process. The idea is based gap and enable the deployment of AV1 codecs. on a statistical analysis performed over a set of VP9 and AV1 bitstreams, which allowed identifying partition modes rarely VP9 owner and developer, Google, is the main company used under certain circumstances. that makes use of VP9 in its video services, like the YouTube platform [11], one of the most popular free video platforms in the world. According to [12] more than 500 hours of video content around the world is published Video Transcoder Bitstream Bitstream every minute on YouTube. All these videos are stored in *.vp9 VP9 AV1 decoder decoded coder *.av1 large data centers and require a huge space in hard drives. VIDEO In [9], [10], and [4] the authors demonstrate that AV1 can achieve a compression efficiency gain of 20%, 18%, Fig. 1. Tandem VP9-to-AV1 transcoder. and 28%, respectively, when

978-9-0827-9705-3 555 EUSIPCO 2020 A. Experimental Setup The Spatial Information and Temporal Information (SI- TI) analysis [15] was performed over all the test sequences available in [16], section “objective-2-slow”, to identify those with most heterogeneous characteristics in order to enable a diverse set of videos to be used for testing. Videos sequences selected for the statistical analysis were Blue Sky, BQ Highway, Dirt Bike, Guitar HDR Amazon, Netflix Dinner Scene, Netflix Food Market 2, Netflix Tunnel Flag, and Water HDR Amazon, as available in [17]. To perform the Fig. 2. Block partitioning allowed in AV1 and VP9 (highlighted). experiments, 60 frames of each video sequence were encoded. The reference codec software for both VP9 and AV1 was used in the experiments. For VP9, the libvpx [18] codec, 128x128 version 1.8.2 (hash code 50d1a4), was used. For AV1, the libaom [5] codec, version 1.0.0 (hash code db8f27), was used. 64x64 The reference software implementations were chosen because 32x32 they represent the most complete versions of the encoder 16x16 specifications, including all the available modes and partitioning possibilities allowed in both formats. 8x8 IETF-NETVC-T [16] is the documentation that defines the 4x4 test configurations used for both video codecs. Following the document, the High Latency CQP configuration was used in (a) (b) the experiments and the Constrained Quality (CQ) parameter was set to values 20, 32, 43, and 55. All experiments are Fig. 3. An AV1 superblock partitioned into blocks: (a) block view, (b) tree view. Gray blocks represent Split partitioning mode. executed sequentially in the same workstation (Intel [email protected] GHz processor, 8 GB RAM, Ubuntu OS), in terminal mode. II. VP9/AV1 PARTITIONING CORRELATION ANALYSIS The CQ parameter was set to 20 for quantization in the A block can assume square or rectangular shapes VP9 encoder, aiming at transcoding from the best quality organized in some configurations, called partitions type. In available in the recommended settings. For AV1, CQ values VP9 there are three partition types: square (named as None), 20, 32, 43, and 55 were used, as defined in [16]. As AV1 vertical (Vert) and horizontal (Horz), as shown in Fig. 2. In introduces new partitioning possibilities in both the horizontal AV1, blocks can assume nine partition types based on the and vertical directions, the occurrence rate of modes three directions observed in VP9. Besides, both codecs also belonging to each direction was summed up in the analysis. allow a Split partition type, which recursively subdivides the Thus, modes 2, 3, 4, and 8 in Fig. 2 are considered as Horz current block into four square blocks. This process follows a modes, whereas modes 1, 5, 6, and 7 are considered as Vert coding block tree structure, as shown in Fig. 3. It is likely that modes. the same block partitioning will be used in both VP9 and AV1 B. Correlation Analysis codecs, since the same video region is being encoded. Considering this, we performed a correlation analysis between For each block in a VP9-encoded video, the same region the partitioning chosen by each codec to use it as a basis for in the AV1-encoded video was observed. For that, a label was the proposed fast transcoding algorithm. attributed to each 4×4 area in the frame, indicating which block size and partition mode were chosen during the

TABLE I. CORRELATION ANALYSIS BETWEEN PARTITION TYPES CHOSEN BY VP9 AND AV1 (CQ 20).

VP9 64×64 32×32 16×16 8×8 AV1 None Horz Vert None Horz Vert None Horz Vert None Horz Vert None 39.18 25.16 10.92 4.56 1.86 1.72 1.61 0.79 1.40 1.49 0.00 0.00 128×128 Horz 1.15 0.38 0.39 0.16 0.06 0.06 0.04 0.12 0.01 0.03 0.00 0.00 Vert 0.07 11.62 0.00 0.03 0.02 0.01 0.19 0.01 0.03 0.01 0.00 0.00 None 36.93 15.36 30.23 12.54 5.02 4.92 3.48 2.36 2.97 5.58 0.00 0.00 64×64 Horz 4.35 16.43 7.34 5.30 3.70 2.16 1.73 1.42 0.93 2.20 0.00 0.00 Vert 6.21 5.25 17.55 5.82 1.78 3.57 1.41 0.90 1.17 2.01 0.00 0.00 None 7.38 13.01 18.84 31.41 19.85 21.65 15.82 12.01 10.40 11.53 0.00 0.00 32×32 Horz 1.62 6.01 4.32 13.32 25.30 13.42 16.45 15.73 9.73 10.97 0.00 0.00 Vert 1.65 2.75 5.48 10.62 10.66 20.63 14.32 7.96 15.67 9.97 0.00 0.00 None 0.83 2.17 2.65 8.52 15.77 16.33 22.78 17.48 21.09 17.19 0.00 0.00 16×16 Horz 0.31 0.85 0.93 4.03 9.53 6.60 10.68 21.50 11.43 16.19 0.00 0.00 Vert 0.24 0.71 0.99 2.79 4.64 7.16 8.82 8.64 18.87 13.23 0.00 0.00 None 0.05 0.21 0.26 0.66 1.40 1.33 2.02 9.33 4.53 6.24 0.00 0.00 8×8 Horz 0.01 0.05 0.05 0.13 0.25 0.21 0.38 1.08 0.67 1.65 0.00 0.00 Vert 0.01 0.05 0.06 0.09 0.14 0.21 0.28 0.46 0.86 1.18 0.00 0.00 None 0.00 0.00 0.01 0.02 0.04 0.04 0.07 0.19 0.22 0.52 0.00 0.00 4×4 Horz 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 Vert 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00

556 test_av1 = None VP9 VP9-to-AV1 Transcoder if size_AV1 == 64x64: bitstream start AV1 encoder . SB encoding if size_VP9 == 64x64 and mode_VP9 == Horz VP9 update modes for testing and cq == 20: frame_info algorithm test_av1 = Horz decoder None Vert frame info mode if size_VP9 == 32x32 and mode_VP9 == Vert ? bsize_AV1 no and cq == 43: decoded < Horz test_av1 = Vert video 8×8 ? search & search & search & encode encode encode yes if size_AV1 == 32x32: None mode Horz modes Vert modes if size_VP9 == 32x32 and mode_VP9 == Horz CQ value and (cq == 20 or cq == 32): stop no split SB subtree ? test_av1 = Horz AV1 split yes if size_VP9 == 16x16 and (mode_VP9 == None bitstream SB subtree or mode_VP9 == Horz) and cq == 20: test_av1 = Horz Fig. 5. Fast transcoding algorithm flow proposed. if size_VP9 == 16x16 and mode_VP9 == Vert and (cq == 20 or cq == 32): Fig. 4 presents the pseudo-algorithm with generalized test_av1 = Vert rules obtained based on the video sequences, CQ values, and if size_AV1 == 16x16: modes, as explained in the analysis presented in section II. if size_VP9 == 16x16 and mode_VP9 == Horz This pseudo-algorithm is based on the block size and the and (cq == 20 or cq == 32 or cq == 43): partitioning mode observed during the VP9 decoding process, test_av1 = Horz and the CQ used. These three parameters determine the occurrence probability of each mode in AV1 re-encoding, Fig. 4. Pseudo-algorithm to select AV1 modes for testing. according to the statistics gathered in the correlation analysis. encoding process for the region. Then, for each video In Fig. 4 test_AV1 represents the selected test mode allowed sequence the occurrences of each combination between block for AV1, size_AV1 represents the current block size and size and partitioning were accumulated and a simple size_VP9 and mode_VP9 are the observed block size and percentage calculation was performed taking the VP9 block partitioning mode observed during the VP9 decoding. size and partitioning combination as anchor. The proposed algorithm, summarized in the Fig. 5, is Table I shows the average correlation between VP9 and implemented in the VP9-to-AV1 transcoder and executed in AV1 partitionings, considering CQ 20. VP9 partitions are three steps. First, the block sizes and partition modes observed presented in the columns, whereas AV1 mode directions are in the VP9 bitstream decoding are exported. During the VP9 presented in the rows. For example, whenever a 32×32 block decoding, the block size, partition type, column and row size and the Horz mode are chosen in VP9, the AV1 encoder positions, and frame number of the superblock is exported to also decided for a 32×32 block and a Horz mode in 25.3% of a file (frame_info). Second, the data is adapted so that the AV1 the cases. Similarly, in 19.9% of the cases, AV1 decided for a codec can handle the values. For example, in VP9 frame 32×32 block and the None mode. It is also noticeable that 4×4 positions are multiples of eight, whereas in AV1 they are blocks are very rarely chosen during the AV1 encoding multiples of four. Third, the AV1 encoder reads frame_info process, even though they are very numerous to be tested and data for every superblock in the re-encoding process. After incur a significant computational cost. that, the pseudo-code to select modes for testing (Fig. 4) is executed before every block size candidate is encoded. If none Table I allows identifying that some partitioning of the allowed partition types result in the best encoding candidates present a much higher probability of occurrence option, AV1 applies the recursive Split process. The Split than others. In general, the higher probabilities are process is allowed from 128×128 to 16×16 blocks only since concentrated around the diagonal area of the table 8×8 blocks are split in less than 0.5% of the occurrences (see (highlighted), which indicates a tendency of AV1 choosing 4×4 blocks in Table I). the same or similar partitions to VP9. The same occurs when other CQ values are used, with a small increase of occurrence IV. EXPERIMENTAL RESULTS for larger blocks in AV1. The full data obtained in the The setup described in section II.A was used for the experiment are available in [17], including the analysis for all experiments presented in this section. The Bjøntegaard Delta CQ values recommended. rate (BD-rate) metric [19] is generally used to compare two video codecs. It represents the bitrate increase (positive value) III. PROPOSED FAST TRANSCODING ALGORITHM or decrease (negative value) considering the same image Based on the statistical analysis presented in the previous quality. In this work, BD-rate values indicate how the fast section, the proposed algorithm follows a strategy that only libaom codec compares to the original libaom codec in terms considers the most likely partitions to occur for every mode of bitrate, considering the same image quality. The Peak observed in VP9 during decoding. Thus, the AV1 encoder Signal-to-Noise Ratio for Luminance (PSNR-Y) and bitrate performs the full encoding loop (i.e., prediction, transform, per second are used to calculate BD-rate. Average time quantization, entropy coding) for only one partition type by savings (TS) are calculated according to (1) considering the block. For example, considering the analysis for CQ 20 four CQ values evaluated, where timefast represents the run (Table I), if a 64×64 block and the Horz mode is detected for a frame region during the VP9 decoding process, then the AV1 encoder should test only the Horz modes when 푡푖푚푒푓푎푠푡 ∑푐푞 ( ⁄ ) 푡푖푚푒표푟푖푔푖푛푎푙 considering a 64×64 block since this is the most probable 푇푆 = 100 ∗ (1 − ) (1) mode class to be chosen for this block size (16.4%). 4

557 TABLE II. COMPRESSION EFFICIENCY AND TIME SAVING RESULTS TABLE III. BD-RATE COMPARISON BETWEEN ORIGINAL AND FAST FOR THE FAST VP9-TO-AV1 TRANSCODER AV1 ENCODERS OVER VP9 ENCODER AS ANCHOR

Class original fast absolute Video Sequence BD-rate (%) TS (%) Ratio Class Video Sequence resolution resolution libaom libaom difference BQFree 4.7555 24.18 0.197 BQFree -17.5447 -13.6765 3.8682 BQZoom 4.4228 23.37 0.189 A A BQZoom -29.6577 -26.9454 2.7123 432x240 Chairlift 4.0765 26.89 0.152 432x240 Chairlift -16.1249 -12.8527 3.2722 Mozzoom 2.7979 26.15 0.107 Mozzoom -20.1364 -17.8323 2.3041 Rain2HDRAmazon 2.9938 32.73 0.091 Rain2HDRAmazon -24.7490 -22,4829 2.2661 RedKayak 1.9096 31.62 0.060 B B RedKayak -0.1282 1.7653 1.8935 640x360 SnowMnt 1.6583 33.61 0.049 640x360 SnowMnt -28.7661 -27.6131 1.1530 TacoManArrows 3.5571 18.65 0.191 TacoManArrows -37.5510 -35.4298 2.1212 Dark 4.1424 23.02 0.180 Dark -38.1851 -36.3775 1.8076 Johnny 6.0483 24.74 0.245 C C Johnny -35.0278 -31.1030 3.9248 1280x720 NetflixDrinvingPOV 6.3701 31.58 0.202 1280x720 NetflixDrinvingPOV -17.2691 -12.3370 4.9321 NetflixRollerCoaster 6.6917 41.73 0.160 NetflixRollerCoaster -25.4314 -20.4854 4.9460 CrowdRun 3.5322 34.25 0.103 CrowdRun -7.3598 -4.2105 3.1493 NetflixCrossWalk 4.0039 18.01 0.222 D D NetflixCrossWalk -19.1816 -16.0049 3.1767 1920x1080 ParkJoy 5.8649 42.67 0.137 1920x1080 ParkJoy -8.5279 -2.0603 6.4676 SeaplaneHDRAmazon 4.9808 24.53 0.203 SeaplaneHDRAmazon -19.6745 -15.7872 3.8873 DOTA2 3.3484 25.49 0.131 DOTA2 -17.5706 -14.9240 2.6466 E Minecraft 4.8084 47.58 0.101 1920x1080 E Minecraft -17.7022 -14.0173 3.6849 4:4:4 Starcraft 2.8454 21.22 0.134 1920x1080 4:4:4 Starcraft -17.6433 -15.4120 2.2313 Wikipedia 8.0104 11.26 0.711 Wikipedia -39.2217 -34.0994 5.1223 Average Group A 4.0132 25.15 0.161 Average Group A -20.8659 -17.8267 3.0392 Average Group B 2.5297 29.15 0.098 Average Group B -22.7986 -20.9401 1.8585 Average Group C 5.8131 30.27 0.197 Average Group C -28.9784 -25.0757 3.9026 Average Group D 4.5955 29.86 0.166 Average Group D -13.6860 -9.5157 4.1702 Average Group E 4.7532 26.39 0.269 Average Group E -23.0345 -19.6132 3.4213 Average ALL 4.3409 28.16 0.178 Average ALL -21.8727 -18.5943 3.2784 Standard Deviation 1.6411 8.92 0.136 Standard Deviation 10.5591 10.8024 1.3405 time of the fast libaom and the timeoriginal is the run time of the original libaom version. Table II shows the experimental results for all video sequences, clustered in five classes. Besides BD-rate and TS for each video, the last column shows the ratio between BD- rate and TS. This ratio indicates the compression efficiency loss for each percent in obtained time savings. Ratios close to zero indicate a better tradeoff between compression efficiency and time savings. Detailed experiment results with data used to calculate both BD-rate and TS can be obtained in [17]. The proposed transcoding algorithm reduces on average 28.16% of the execution time, at the cost of an average BD- rate of 4.3409%. TS results present a standard deviation of Fig. 6. Comparison between VP9, original libaom and proposed fast 8.92%, whereas BD-rate values present a standard deviation libaom for the Snow Mnt sequence. of 1.6411%. The best-case shown in Table II corresponds to the Snow Mmt video, Group B. This video presents a TS of 33.61% at the cost of a BD-rate increase of 1.6583%. On the other hand, the worst case is the Wikipedia video, Group E, which achieves a TS of 11.26% at the cost of a BD-rate increase of 8.0104% in comparison to the original libaom implementation. Wikipedia is the only screen content video used in the experiments. As AV1 includes specific techniques to encode this type of media, the proposed algorithm which was devised for general-content sequences, is not able to take advantage of this specific case. As previously mentioned, the current AV1 version achieves an average BD-rate decrease of 22% over VP9. This is noticeable in Table III, which shows a comparison between Fig. 7. Comparison between VP9, original libaom and proposed fast the original libaom and the proposed fast libaom version, libaom for the Wikipedia sequence. taking the VP9 encoder as the baseline. For that, the BD-rate values were calculated using the same input video sequences libaom version for transcoding achieves a BD-rate decrease of for both VP9, original libaom and fast libaom software. When 18.5% in comparison to VP9, which is a slight difference compared to VP9, the original AV1 encoder achieves an given the computational cost decrease in only 3.2% of BD- average BD-rate decrease of 21.8%, as expected. The fast Rate. These results show that even when employing the fast

558 libaom version for transcoding, AV1 still excels by far in [4] Y. Chen et al., “An Overview of Core Coding Tools in the AV1 Video comparison to VP9 in terms of compression efficiency. Codec,” in 2018 Picture Coding Symposium PCS 2018 - Proceedings, pp. 41–45, 2018. doi: 10.1109/PCS.2018.8456249 Rate-distortion efficiency results are presented in the [5] “AV1 Codec Library,” Alliance for Open Media, 2015. [Online]. charts of Fig. 6 and Fig. 7 for the best and worst cases Available: https://aomedia.googlesource.com/aom. [Accessed: 20- presented in Table II (Snow Mnt and Wikipedia, respectively). Jan- 2020]. The figures compare the compression efficiency between the [6] “Scalable Video Technology for AV1 Encoder,” Intel, 2015. [Online]. Available: https://github.com/OpenVisualCloud/SVT-AV1. video sequences encoded with VP9, the original libaom, and [Accessed: 20- Jan- 2020]. the proposed fast libaom encoder. Notice that in both the best [7] T. Davies, “Cisco Leap Frogs H.264 Video Collaboration with Real- and worst-case scenarios, the proposed encoder presents Time AV1 Codec,” CISCO, 2019. [Online]. Available: compression efficiency results very close to the original https://blogs.cisco.com/collaboration/cisco-leap-frogs-h-264-video- libaom. Also, it is clearly noticeable that AV1 presents a much collaboration-with-real-time-av1-codec. [Accessed: 20- Jan- 2020]. better compression efficiency than VP9, either the original [8] “rav1e,” Xiph.org 2015. [Online]. Available: video transcoder or the proposed fast video transcoder. https://github.com/xiph/rav1e. [Accessed: 20- Jan- 2020]. [9] D. Grois, T. Nguyen, D. Marpe, “Performance comparison of AV1, V. CONCLUSION JEM, VP9, and HEVC encoders,” in SPIE 10396, Applications of Digital Image Processing XL, 103960L. Proceedings, 2018. doi: Launched recently by the libaom consortium, AV1 will 10.1117/12.2283428 gradually become a widely adopted encoding format for [10] L. Guo, J. De Cock and A. Aaron, “Compression Performance online video streaming. Big companies with a large amount of Comparison of x264, x265, libvpx and aomenc for On-Demand video content will eventually migrate from previous formats Adaptive Streaming Applications,” in 2018 Picture Coding to AV1. Thus, transcoding from VP9 to AV1 is in the interest Symposium (PCS), San Francisco, CA, 2018, pp. 26-30. doi: 10.1109/PCS.2018.8456302 of several companies that employ royalty-free formats. [11] M. Srinivasan, “VP9 Encoder and Decoders for Next Generation However, fast transcoding approaches are required due to the Online Video Platforms and Services,” in SMPTE 2016 Annual high computational cost required by the AV1 encoding Technical Conference and Exhibition, Los Angeles, CA, 2016, pp. 1- process. This paper presents an algorithm that allows reusing 14. doi: 10.5594/M001734 block partitioning decisions provided by the VP9 decoding [12] J. Clement, “Hours of video uploaded to YouTube every minute,” process to accelerate the VP9-to-AV1 transcoding. The Statista, 2019. [Online]. Available: strategy is based on a statistical analysis that allowed https://www.statista.com/statistics/259477/hours-of-video-uploaded- identifying the most probable partitions in AV1 when certain to-youtube-every-minute/. [Accessed: 20- Jan- 2020]. [13] A. J. Díaz-Honrubia, J. L. Martínez, P. Cuenca, J. A. Gamez and J. M. modes and blocks sizes are observed during VP9 decoding. Puerta, “Adaptive Fast Quadtree Level Decision Algorithm for H.264 By doing so, the libaom execution time is reduced by 28.16% to HEVC Video Transcoding,” in IEEE Transactions on Circuits and on average, at the cost of a BD-rate increase of 4.3409%. Also, Systems for Video Technology, vol. 26, no. 1, pp. 154-168, Jan. 2016. when compared to the VP9 compression efficiency, the fast doi: 10.1109/TCSVT.2015.2473299 AOM encoder still yields a BD-rate gain of 18.6%, which is [14] A. Borges, B. Zatt, M. Porto and G. Correa, “Fast Hevc-to-Av1 close to the obtained by the original AOM (21.8%). Transcoding Based On Coding Unit Depth Inheritance,” in 2019 IEEE International Conference on Image Processing (ICIP), Taipei, Taiwan, 2019, pp. 3571-3575. doi: 10.1109/ICIP.2019.8803482 ACKNOWLEDGMENTS [15] International Telecommunication Union, “Recommendation ITU-T This study was financed in part by the Coordenação de P.910: Subjective Video Quality Assessment Methods for Multi-Media Aperfeiçoamento de Pessoal de Nível Superior - Brasil Applications,” ITU-T, 2008. (CAPES) - Finance Code 001, the Conselho Nacional de [16] T. Daede, A. Norkin, I. Brailovskiy, “Video Codec Testing and Quality Desenvolvimento Científico e Tecnológico (CNPq), and the Measurement draft-ietf-netvc-testing-08,” Internet Engineering Task Fundação de Amparo à Pesquisa do Estado do Rio Grande do Force, 2019. [Online]. Available: https://tools.ietf.org/html/draft-ietf- netvc-testing-08. [Accessed: 20- Jan- 2020]. Sul, Brasil (FAPERGS). [17] A. Borges, “EUSIPCO 2020,” Github, 2020. [Online]. Available: https://github.com/amborges/EUSIPCO_2020. [Accessed: EFERENCES R 28/02/2020]. [18] “WebM VP8/VP9 Codec SDK,” Chromium, 2019. [Online]. [1] D. Mukherjee et al., “A Technical Overview of VP9—The Latest Available: https://chromium.googlesource.com/webm/libvpx/. Open-Source Video Codec,” in SMPTE Motion Imaging Journal, vol. [Accessed: 20- Jan- 2019]. 124, no. 1, pp. 44-54, Jan. 2015. doi: 10.5594/j18499 [19] G. Bjøntegaard, “VCEG-M33: Calculation of Average PSNR [2] “Daala video compression,” Xiph.org, 2013. [Online]. Available: Differences between RD curves,” Video Coding Experts Group https://xiph.org/daala/. [Accessed: 20- Jan- 2020]. (VCEG) 2001. [3] G. Bjøntegaard, T. Davies, A. Fuldseth and S. Midtskogen, “The Thor Video Codec,” in 2016 Data Compression Conference (DCC), Snowbird, UT, pp. 476-485, 2016. doi: 10.1109/DCC.2016.74

559